Pollen Robotics, the Bordeaux-based robotics company acquired by Hugging Face, launched Microduck on August 27 — a 25cm tall biped robot with 15 motors, a camera, LiDAR, and an articulated beak, priced at $399 and designed to make reinforcement learning something you do on a desk rather than in a robotics lab. Pre-orders are open now, with first deliveries targeted before Christmas 2026 in North America, Europe, and the UK.
What Microduck Actually Does
This is not a toy. It is a research platform small enough to fail safely. Microduck ships with seven learned behaviours: walking, sitting and standing, kicking, grabbing objects with its beak, getting back up from a fall, and roller skating. Each behaviour is a reinforcement learning policy trained in simulation and deployed on the physical robot using sim-to-real transfer.
The robot runs a 50Hz onboard policy loop. It has 15 motors, a front-facing camera, a small depth sensor (LiDAR), and two IMUs. It weighs under 800 grams. The beak is functional — it dips, scoops, and carries small objects. The Register noted that Pollen demonstrated it picking up socks and a marker, placing them in a box.
What makes this different from a programmable toy is the training pipeline. The full RL stack — robot control, simulation, training, and sim-to-real deployment — is open source on GitHub under Apache 2.0. Behaviours are trained in MuJoCo physics simulation, then transferred to the real robot in one step. You can retrain any shipped behaviour on your own machine, or on Hugging Face Jobs, and deploy it back to the robot.
Why Small Matters
The honest argument for making a robot this small is about failure. On a full-size humanoid, every failed training attempt can be expensive, difficult to reset, or dangerous. On Microduck, a failed behaviour ends with a small robot on the floor. It self-recovers from many common falls. You do not have to pick it up after every bad attempt.
That changes where learning can happen. A classroom, a bedroom, a workbench — anywhere with a laptop and a flat surface. The price matters here too. A $399 robot that trains policies is a different proposition from a $16,000 research platform. Pollen’s previous robot, Reachy Mini, sold over 10,000 units. Microduck is aimed at the same audience: people who want to learn by doing, not by reading documentation.
The Hugging Face Connection
The Register’s coverage described Pollen as “a Hugging Face subsidiary,” which is the connection worth paying attention to. Hugging Face is the platform where most open-source AI models live. It is also the company Nvidia agreed to buy for $12.9 billion this week.
Pollen’s blog post made the Hugging Face fit explicit: “Models are useful because people can build on one another’s work. We want physical behaviours to become easier to share in the same way, together with the environments and training recipes needed to reproduce them.”
The implication is that robot behaviours could become shareable artifacts — train a walking gait, publish the policy, someone else downloads it and deploys it on their Microduck. That is the model that made Hugging Face central to AI model distribution, applied to physical AI. Whether it works for robotics the way it worked for language models is an open question, but the infrastructure is now in place to try.
Open Source, For Real
The software stack is genuinely open. The main repository contains the SDK and robot software. The RL repository contains the reinforcement learning and sim-to-real tools. MuJoCo, the physics simulator, is open source. The licence is Apache 2.0, which means you can read, fork, modify, and commercially reuse the code.
This is the kind of open-source robotics approach we wrote about in our New Zealand number 8 wire piece — where the barrier to entry is not the cost of hardware but the willingness to learn. A $399 robot with a full training stack and no licensing restrictions is about as low as that barrier gets.
What It Cannot Do
Microduck will not clean your house. It cannot carry heavy objects. It does not have voice interaction or conversational AI — that is Reachy Mini’s territory. The two robots share a philosophy but serve different purposes. Reachy Mini is a platform for AI that interacts. Microduck is a platform for AI that acts.
The behaviours shipping today are a starting point, not a final state. Pollen was clear about this: “Microduck will not solve every hard problem in robotics, and that is not the goal.” The goal is to make those problems easier to approach, and to give more people a practical way to learn how AI moves from simulation into a real machine.
A Kiwi Angle
For New Zealand, the interesting question is whether a $399 reinforcement learning platform finds its way into classrooms and polytechs. NZ’s digital technologies curriculum already includes computational thinking and designing digital outcomes. A robot that costs less than a graphics card, runs open-source software, and lets students train physical AI behaviours from scratch is the kind of tool that could make robotics education tangible — without a lab or a budget application.
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
